The AI log, science · open weights
14 Nov 2026
cuBLAS Ozaki-II, cuPhoton xBeam, Unsloth Dynamic V3.1, Nex-N2.6
NVIDIA speeds up FP64 matrix math and black hole correlation, Unsloth refreshes Qwen3.8-27B quants, and Nex AGI posts open Nex-N2.6 weights.
By Ali Asif, published
What shipped
- NVIDIA
Faster Scientific Simulations with New FP64 Algorithms in NVIDIA cuBLAS
CUDA Toolkit 13.4.2 adds Ozaki-II and 2M emulation that run double-precision matrix multiplies on integer Tensor Cores, with DGEMM reported 5.5x faster than native FP64 on VR NVL72.
- NVIDIA
Run GPU-Powered Black Hole Interferometry with NVIDIA cuPhoton
The xBeam correlator turned a 10 TB Event Horizon Telescope dataset into visibilities in 70 seconds on 64 GPUs, against 31,992 seconds on an x86 CPU baseline.
- Unsloth
Qwen3.8 - How to Run Locally (Dynamic V3.1 update)
New Qwen3.8-27B GGUFs use Dynamic V3.1, which Unsloth says improves average accuracy by 5 to 10 percent at the same size across its benchmark set.
- Nex AGI
nex-agi/Nex-N2.6
An Apache-2.0 image-text-to-text model with about 321 billion parameters, stored mostly in FP8, loadable with Transformers, vLLM and SGLang.
A quieter day for launches, but two NVIDIA engineering posts and two open-weights items are worth a look. Every number below comes from the pages I opened. I have not run any of it.
cuBLAS gets Ozaki-II and 2M emulation
CUDA Toolkit 13.4.2 adds two algorithms that do double-precision matrix multiplication on integer Tensor Cores while keeping full FP64 accuracy. Ozaki-II cuts the number of integer multiplications compared with Ozaki-I by using the Chinese Remainder Theorem. The 2M method turns a complex matrix multiply from four real multiplications into two.
The reported speedups over native FP64 are specific. DGEMM with Ozaki-II runs 5.5x faster on VR NVL72 and 3.9x faster on GB200. ZGEMM with 2M is 5x faster on GB200. In quantum chemistry CCSD runs, end-to-end speedups are 2x on VR NVL72 and 1.4x on GB200 at full FP64, rising to 5x and 3.5x in mixed precision. BerkeleyGW materials science runs gain 6.6x and 3.9x. The post also notes that Rubin GPUs add a TI16 numeric format and a tcgen05.mma instruction that speed these methods further.
Developers control it through cublasSetEmulationStrategy and manual mantissa bit settings, and NVIDIA points to cuBLAS emulation samples on GitHub.
cuPhoton xBeam for the Event Horizon Telescope
The second NVIDIA post covers xBeam, a GPU correlator in cuPhoton. It reads VDIF voltage recordings, applies delay and fringe corrections, channelizes with FFTs, cross-multiplies station pairs and writes visibilities in DiFX SWIN or xbeam-v1 format.
On a Vera Rubin NVL72 system with 64 GPUs and GPUDirect Storage, a 10 TB dataset took 70 seconds, a 460x speedup over a 31,992 second x86 CPU baseline. Work splits across GPUs on one node or across nodes with shard flags, and a merge step combines partial outputs. The post also describes AI agents orchestrating the pipeline by reading VDIF metadata, writing JSON job specs and launching runs.
Unsloth refreshes Qwen3.8-27B quants
Unsloth released new Qwen3.8-27B GGUFs on November 11 using Dynamic V3.1, a preview of its updated quantization algorithm. It says average accuracy improves 5 to 10 percent at comparable sizes, tested on Terminal Bench 2.1, DeepSWE v1.1, LiveCodeBench v6, GPQA Diamond, MMLU-Pro and AIME25. Agentic tests used 256K context and the rest 128K. Most quants from 2-bit to 6-bit were updated, while 3-bit stayed because it matched V3.1. The model is a 27B dense model with vision, reasoning and a 256K context window that Unsloth says runs on 17GB of RAM or VRAM.
Nex-N2.6 weights
Nex AGI posted Nex-N2.6 on Hugging Face under Apache-2.0. The metadata lists roughly 321 billion parameters, mostly in FP8, a glm5_next architecture and an image-text-to-text pipeline. The card gives usage snippets for Transformers, vLLM and SGLang but no benchmark table, so I will not guess at quality.
What I'd watch next
Whether integer-based FP64 emulation shows up in open scientific codes outside NVIDIA's own samples, independent checks of Dynamic V3.1 on coding tasks, and whether Nex AGI publishes evaluations for N2.6.
Sources
Topics: NVIDIA, Qwen, Open Weights, Nex AGI